Towards Human-like Empathy: Confidence-Driven Fast-Slow Reasoning for Multimodal Empathetic Response Generation
Abstract
Multimodal Empathetic Response Generation (MERG) integrates textual, visual, and acoustic signals to synthesize nuanced affective responses. While Large Language Models (LLMs) demonstrate strong capabilities in emotional reasoning, they still suffer from reasoning redundancy and contextual prolixity. Applying static, heavy-duty reasoning to all queries wastes computational resources, while uncompressed multimodal histories introduce excessive and noisy tokens into LLM inputs. To address these challenges, we propose the CoRE framework. Inspired by dual-process cognition, CoRE introduces a confidence-driven fast-slow reasoning mechanism that dynamically toggles between a lightweight fast path for intuitive fluency and an expanded slow reasoning chain for resolving complex emotional ambiguity. Additionally, a condensed-temporal dialogue encoder module distills noisy historical signals into compact representations, reducing the multimodal token burden. Extensive experiments on the large-scale AvaMERG benchmark demonstrate that CoRE achieves competitive performance, outperforming existing LLM baselines in emotion recognition, semantic alignment and human-evaluated empathy while optimizing inference efficiency. The source code is available at https://github.com/Cathryn1103/CoRE_MERG.